DUALFloodGNN: Physics-informed Graph Neural Network for Operational Flood Modeling

AI summary

Overview

Research area: Physics-informed machine learning, specifically graph neural networks (GNNs) applied to hydrodynamic flood modeling and operational flood forecasting.

Technical level: Advanced. The paper assumes familiarity with graph message passing, physics-informed loss design, autoregressive rollout training, and shallow-water mass conservation concepts.

Scope: The paper introduces DUALFloodGNN, a GNN architecture that jointly predicts node water volume and edge water flow while enforcing mass conservation at both global and local scales, and benchmarks it against standard GNNs and existing physics-informed flood GNNs.

What This Paper Is About

Conventional physics-based numerical flood models solve the governing fluid-flow equations accurately but are too computationally expensive for real-time or operational forecasting. Data-driven deep learning surrogates are much faster but are often treated as "black boxes" that are not guaranteed to respect physical laws. The paper's goal is to build a fast, physically consistent flood surrogate that explicitly embeds mass conservation into the learning objective rather than assuming it implicitly.

Key Contributions

1. A dual-scale physics-informed loss function that promotes mass conservation at both the global (whole-catchment) and local (individual node) scales through separate regularization terms, so that conservation is preserved within localized areas of the computational mesh rather than only in aggregate.
2. A novel architecture that predicts node and edge features simultaneously. DUALFloodGNN maps water flow as an edge attribute to preserve its vector character, and uses a shared message-passing operation to generate messages that update both node embeddings (water volume) and edge embeddings (water flow), using only the original graph structure — avoiding the line-graph construction required by domain-agnostic node-and-edge GNNs such as CensNet and NENN.
3. A dynamic curriculum learning training strategy that replaces the fixed rollout-length schedule of prior work with an early-stopping mechanism, extends the prediction horizon only after validation performance converges at the current rollout length, uses a tunable step size hyperparameter (C), and decays the learning rate by a factor (gamma) after each curriculum step.
4. The first benchmark across different GNN-based flood models, comparing DUALFloodGNN against standard GNN architectures (GCN, GAT, GraphSAGE, GIN, GINE) and domain-specific models (HydroGraphNet, SWE-GNN) for water volume, flow, and depth prediction.

Main Findings

• Best node water volume accuracy: DUALFloodGNN achieved an RMSE of 2,216.27 ± 868.21 m³ and MAE of 1,075.63 ± 394.83 m³, improving on the second-best model (GINE, RMSE 3,437.68 ± 693.94 m³) by 35.53%. Other baselines were GCN (9,849.98 ± 2,485.73 m³), GAT (13,232.86 ± 11,835.18 m³), GraphSAGE (11,146.97 ± 11,420.79 m³), GIN (5,023.18 ± 1,450.37 m³), and HydroGraphNet (7,584.51 ± 3,068.54 m³).

• Best edge water flow accuracy: DUALFloodGNN reached an RMSE of 25.88 ± 9.64 m³/s and MAE of 11.40 ± 4.17 m³/s, a 41.21% improvement over the baseline RMSE. Baselines included GCN (90.99 ± 25.04 m³/s), GraphSAGE (44.02 ± 12.54 m³/s), GIN (51.21 ± 16.33 m³/s), GINE (249.31 ± 70.59 m³/s), and GAT (263.40 ± 63.16 m³/s).

• Edge-feature baselines underperform without joint modeling: The paper reports that baseline models incorporating edge attributes (GAT and GINE) performed significantly worse in flow prediction, which the authors interpret as evidence that joint node-and-edge modeling is necessary to exploit edge features.

• Single model for both targets: Only DUALFloodGNN produced metrics for both water volume and flow within one model; the other baseline architectures required separate models for each prediction task.

• Best water depth accuracy: With depth back-calculated from predicted volume using a volume-to-elevation mapping from the HEC-RAS simulation, DUALFloodGNN achieved RMSE 0.21 ± 0.07 m and MAE 0.07 ± 0.03 m, a 72.37% improvement over the second-best RMSE. HydroGraphNet scored RMSE 0.76 ± 0.34 m and MAE 0.29 ± 0.16 m; SWE-GNN scored RMSE 1.23 ± 0.35 m and MAE 0.61 ± 0.21 m.

• Flood extent classification: DUALFloodGNN attained CSI of 0.90 ± 0.03 at the 0.05 m threshold and 0.91 ± 0.03 at the 0.3 m threshold, i.e. similar accuracy at both thresholds. HydroGraphNet attained 0.69 ± 0.13 and 0.80 ± 0.05, showing a high delta between thresholds; SWE-GNN attained 0.47 ± 0.11 and 0.49 ± 0.09.

• Inference speed: DUALFloodGNN's inference time was 4.10 ± 0.06 s, close to HydroGraphNet's 3.25 ± 0.17 s (a 0.85 second difference) and faster than SWE-GNN's 7.21 ± 0.20 s. The authors state this preserves the 2-to-3 orders of magnitude speed-up over numerical solvers reported in prior studies, giving the example that an event taking almost a day for a numerical model takes around 3 minutes with their model.

• Temporal stability and spatial fidelity: Across all three tasks DUALFloodGNN maintained the lowest per-timestep RMSE throughout the rollout period. In the maximum water depth maps, SWE-GNN predicted mostly uniform flood propagation across the catchment and HydroGraphNet overestimated flooding at a few nodes far from the central river area, while DUALFloodGNN produced the most accurate map.

• Inflow feature dominates performance: Removing the global inflow feature degraded node RMSE to 20,562.71 ± 8,054.11 m³ and edge RMSE to 423.40 ± 246.11 m³/s. HydroGraphNet showed the same degradation when its inflow feature was omitted (node RMSE 20,867.39 ± 6697.08 m³). The authors attribute this to the flow-dominant nature of the events in the dataset, and enhance the feature by treating it as dynamic and including previous boundary conditions.

• Local physics loss matters most: Ablations gave node/edge RMSE of 2,794.10 ± 851.61 m³ and 48.09 ± 14.52 m³/s for global-loss-only; 2,116.45 ± 845.44 m³ and 26.07 ± 8.55 m³/s for local-loss-only; and 2,640.42 ± 1011.80 m³ and 46.86 ± 15.81 m³/s for no physics loss. Local regularization alone outperformed the purely data-driven model and the global-only variant, while the combined global-plus-local model performed comparably to local-only, with the better variant varying across testing folds.

Methodology in Plain English

The study targets a section of the Wollombi River watershed in New South Wales, Australia. Using the HEC-RAS numerical model, the authors generated a dataset of 56 flow-dominant flood events. The 2D unstructured mesh was converted into a directed graph of 1129 nodes and 2743 edges, where each mesh cell centroid is a node and connections between adjacent cell centers are edges. Each event spans 576 time steps at 15-minute resolution, and models were evaluated with 14-fold cross validation.

The architecture follows an encode-process-decode pattern. A separate encoder transforms node features (static properties such as area and elevation, plus dynamic quantities such as volume, flow, rainfall, and global inflow over the previous p timesteps) and edge features into a shared latent space. Inside the processor, messages are generated for each node from itself, its neighbors, and the connecting edge. Critically, the same messages update both node embeddings (summed over neighbors and passed through a fully connected layer) and edge embeddings (overwritten directly by the message), with residual connections at each update.

Rather than predicting future states directly, the model predicts the change in water volume and water flow from one timestep to the next, which is added to the current state as a residual connection — reducing output variation and stabilising long rollouts.

Two loss terms are combined. The prediction loss is a weighted sum of mean squared error on node volume change and edge flow change. The physics loss adds absolute-value residuals for mass balance: the global term compares total predicted volume change against boundary inflow minus outflow plus rainfall across the entire catchment, while the local term does the same for each individual node, using edge flow predictions to compute each node's inflow and outflow. Local conservation therefore couples node and edge outputs, whereas global conservation constrains only node volume because inflow, outflow, and rainfall are supplied as boundary conditions.

Training uses a multi-step-ahead rollout loss: for each batch the model runs autoregressively for O steps and the losses are averaged, forcing the model to become robust to its own noisy predictions and to conserve mass along the entire trajectory. The curriculum schedule from prior work is made adaptive — the horizon is extended only when validation performance stops improving at the current length, with a tunable step size and learning-rate decay between stages.

Why This Matters

Impact on research. The paper shows that explicitly enforcing mass conservation at the local level — not just globally — improves both regression accuracy and flood-extent classification over a purely data-driven model and over a global-only physics model. It also argues that joint node-and-edge modeling with shared messages is necessary for edge attributes to help rather than hurt, and it provides the first cross-model benchmark of GNN flood architectures. The code and dataset are open sourced (github.com/acostacos/dual_flood_gnn; dataset at hdl.handle.net/2123/35293, DOI 10.25910/9xav-0s86).

Real-world applications.

• Operational flood forecasting, where a simulation that would take a numerical solver almost a day can be produced in around 3 minutes.
• Flood extent mapping and evacuation planning, supported by CSI of 0.90 and 0.91 at the 0.05 m and 0.3 m depth thresholds.
• Flood-resilient engineering design, where rapid scenario simulation can inform infrastructure decisions.
• Disaster management and strategic planning, where multiple hydrologic variables (volume, flow, depth) are needed from a single model.

Industry relevance. Insurance, civil infrastructure, and emergency-management sectors depend on fast, physically credible flood information. The reported 2-to-3 orders of magnitude speed-up over numerical solvers, combined with inference times comparable to the fastest domain-specific baseline, positions the approach as a candidate for operational deployment — though the authors present results for a single catchment and do not report wall-clock training costs or the numeric values of the loss weighting coefficients.

Future Directions

• Cross-catchment and cross-flood-type generalization. The authors explicitly propose evaluating the model on other catchments and flood types, since the current study covers one watershed section in New South Wales, Australia.
• Non-uniform density conditions. In flows with high sediment or debris concentration, density is not spatially uniform, which the authors note may affect the volume-based regularization that assumes constant density to equate mass and volume conservation.
• Choosing between local-only and combined physics losses. Because the local-only and global-plus-local variants perform comparably and the better one varies across testing folds, the authors state that selection should be driven by empirical evaluation on the target dataset — leaving open the question of a principled rule for that choice.
• Tuning of the physics and prediction loss weights. The coefficients weighting the node, edge, global, and local loss terms are presented symbolically; their specific values and the resulting sensitivity are not reported in the content.

Target Audience

This paper is most useful to machine learning researchers working on physics-informed neural networks and graph learning, and to hydrologists, hydraulic engineers, and flood-risk practitioners evaluating deep learning surrogates for operational forecasting. It will also interest researchers in scientific machine learning who care about embedding conservation laws into autoregressive, multi-step prediction pipelines, and engineering teams assessing whether GNN surrogates can replace or augment expensive numerical solvers in time-critical settings.

Authors’ abstract

Flood models inform strategic disaster management by simulating the spatiotemporal hydrodynamics of flooding. While physics-based numerical flood models are accurate, their substantial computational cost limits their use in operational settings where rapid predictions are essential. Models designed with graph neural networks (GNNs) provide both speed and accuracy while having the ability to process unstructured spatial domains. Given its flexible input and architecture, GNNs can be leveraged alongside physics-informed techniques with ease, significantly improving interpretability and generalizability. We introduce a novel flood GNN architecture, DUALFloodGNN, which embeds physical constraints at both global and local scales through explicit loss terms. The model jointly predicts water volume at nodes and flow along edges through a shared message-passing framework. To improve performance for autoregressive inference, model training is conducted with a multi-step loss enhanced with dynamic curriculum learning. Compared with standard GNN architectures and state-of-the-art GNN flood models, DUALFloodGNN achieves substantial improvements in predicting multiple hydrologic variables (e.g., water volume, flow, and depth) while maintaining high computational efficiency. The model is open sourced at https://github.com/acostacos/dual_flood_gnn. The dataset is open sourced at https://hdl.handle.net/2123/35293 with the DOI 10.25910/9xav-0s86.